An address change doesn’t need a reasoning engine. The customer gives a new address, a flow validates it, the record updates, and a confirmation goes out. That path was the same in 2020 and it’ll be the same next year.
I keep that request in mind when I read about Agentforce, which went generally available last month at $2 a conversation. The launch material makes it sound like a new kind of system. Once I took it apart I counted five parts, and one of them is automation most Salesforce orgs already have.
Five parts, in the order a request meets them
The first part is a classifier. The customer types something (the utterance, in Salesforce’s wording), and the first model call only has to decide which topic the message belongs to, out of a list you wrote. A topic is a named job with a description, like “coverage questions” or “appointment changes”. If nothing fits, the agent says so or hands off.
Second, instructions scoped to that topic. Each topic carries its own plain-language rules, such as what to ask for and what never to promise. The scoping is the design decision I like most here. The model sees only the rules for the job it’s doing, which keeps the prompt short and the behavior easier to test.
Third come the actions, and this is the part you already own. An action is an autolaunched flow, an invocable Apex method, or a prompt template from Prompt Builder, each wrapped with a description the model can read. The agent can’t do anything in your org that isn’t on a topic’s action list.
Fourth is grounding data: records from the org, knowledge articles, and anything you’ve unified in Data Cloud, which is a second store with its own clocks. It gets retrieved and placed in the prompt so the answer comes from your data and not from the model’s memory of warranties in general.
Last are the guardrails. The Einstein Trust Layer sits around the model calls, and it’s where data masking, toxicity scoring and the audit trail live. Check which of those are switched on for agents in your own org before you count on them, because the defaults differ by feature. Which fields are allowed into a prompt at all is a separate decision, and I wrote about it in which CRM fields may reach a model.
In the middle is a loop, which Salesforce calls the Atlas Reasoning Engine. It reads the instructions and the data, picks an action, looks at the result, and decides whether to pick another or answer. That loop is the one piece that’s new to a Salesforce org. The five parts are configuration around it.
What the loop buys you
A flow needs every branch drawn before the first customer arrives. That’s its strength. You can read it and test each path, so you know what it will do on a Tuesday night.
The loop is for requests where you can’t draw the branches. The customer writes “the thing under my sink is making a noise again, is that covered or do I have to pay like last time”. Nobody can build a decision element for that sentence. A model can work out that it’s a coverage question about a garbage disposal with a prior claim, look up the contract, and answer.
My test is a whiteboard. If I can draw the request as a flowchart, it stays a flow. Paying a model to walk a path that’s already known adds cost and latency, and every so often the model will pick a different path.
Flow Orchestration falls on the flow side of that line too. A process with five stages and three approvers is long, but every step is known in advance. Long and known is a job for a workflow that remembers where it is.
Ten request types, three verdicts
Hearthguard Home Warranty is a company I’ve invented, and so is its queue: 14,000 cases a month, with the ten most common request types covering 82 percent of them, about 11,480 cases.
Six of those types follow fixed paths: address changes, invoice copies, and four more like them. Together they’re roughly 7,100 cases. Three types need someone to read free text and look something up, around 3,300 cases. The tenth is claim disputes, about 1,080.
I put a few of them to three questions:
| Request type | Path known in advance | Inputs structured | Cost of a wrong action | Verdict |
|---|---|---|---|---|
| Address change | Yes | Yes | Low, easy to reverse | Flow |
| Invoice copy | Yes | Yes | Low | Flow |
| Coverage question | No | No, free text | Medium, sets an expectation | Agent, read-only actions |
| Claim status, described loosely | No | No | Low | Agent, read-only actions |
| Goodwill credit under $50 | No | Mixed | Medium, real money | Agent with approval |
| Claim dispute | No | No | High, money and a contract | Person |
The goodwill credit row isn’t in Hearthguard’s top ten. I added it because none of the ten lands on the in-between verdict, where the agent prepares the action and a person approves it before anything is written.
Now the money. Sending the six fixed types through the agent means about 7,100 conversations at $2 each, which is $14,200 a month for work the flows already finish correctly. The three agent topics cost about $6,600, and that buys something the flows couldn’t do.
Claim disputes go to a person. The path is unknown and a wrong answer is expensive, and I wouldn’t put a model on that combination.
A one-page spec per topic
Before anyone opens Agent Builder, I’d want each topic written down in a form the support lead can read and argue with. The format is my own planning document, so nothing in Salesforce will import it, and the action names are placeholders.
topic: Coverage questions
owner: Support operations lead
scope:
- Answer whether an appliance or system is covered under the customer's active contract
- Explain the service fee and coverage limit for that item
out_of_scope:
- Approving, denying, or estimating the payout of a claim
- Changing a contract, an address, or a payment method
- Any question about a contract the customer is not verified on
actions_allowed:
- name: Get Active Contract
type: flow
access: read
- name: Get Coverage Terms For Item
type: flow
access: read
- name: Summarize Coverage Clause
type: prompt_template
access: read
grounding:
- Contract and contract line records
- Published coverage knowledge articles
escalation_rule:
- Customer mentions a dispute, a denial, or a refund
- Contract lookup returns no match or more than one
- Customer asks for a person
- Two turns pass without a grounded answer
review:
sample: 50 conversations a week
reviewer: Support quality analyst
The out-of-scope list does more work than the scope list. Most of the instructions you’ll write in the builder come from it. And if every action on the page is read-only, the worst the topic can do is give a wrong answer, which the weekly sample is there to catch.
Write one of these for each row that came out as “agent”. At Hearthguard that’s three specs, and the $14,200 a month that would have gone on sending the six fixed types through a model stays available for the topics that need one.